PriRoAgg: Achieving Robust Model Aggregation With Minimum Privacy Leakage for Federated Learning

Sizai Hou, Songze Li, Tayyebeh Jahani-Nezhad, Giuseppe Caire

TL;DR

PriRoAgg is a novel framework that solves the "privacy vs. robustness" conflict in Federated Learning by introducing the concept of Aggregated Privacy. It uses Lagrange Coded Computing and distributed zero-knowledge proofs to execute advanced robust aggregation algorithms while revealing only the bare minimum statistical summaries needed to filter out malicious updates.

What Problem Does This Paper Address?

Federated Learning (FL) is built on a paradox. To protect privacy, we want to hide individual user updates from the central server. However, to ensure robustness, the server needs to inspect those updates to identify "Byzantine" (malicious) users who might be trying to poison the model or plant backdoors.

Existing solutions usually fail in one of two ways:

  1. Weak Security: They use simple checks (like limiting the "norm" or size of an update) that advanced attackers can easily bypass.
  2. Privacy Leaks: They use complex algorithms (like multi-Krum) that inadvertently reveal sensitive information, such as the "distance" between different users' private data.

The authors address this by asking: What is the absolute minimum amount of information a server needs to perform a robust aggregation without compromising individual privacy?

Methods

The authors introduce PriRoAgg, a framework built on three sophisticated pillars:

  • Aggregated Privacy: A new formal security definition. Instead of trying to hide everything, it allows the server to see the sum of certain functions of user data (e.g., the sum of the signs of the gradients) but absolutely nothing about any individual user.
  • Lagrange Coded Computing (LCC): A specialized form of secret sharing. It allows users to split their data into "shares" such that the server can perform linear computations on the encrypted pieces. It is significantly more efficient than standard Homomorphic Encryption.
  • Distributed Zero-Knowledge Proofs (SNIP): To prevent malicious users from lying about their secret shares, PriRoAgg uses an optimized version of SNIP. This allows users to prove they followed the rules (e.g., "I correctly calculated the sign of my gradient") without revealing the gradient itself.

The paper instantiates this framework into two specific protocols:

  • PriRoAgg-RLR: Uses sign-based information to defend against backdoor attacks.
  • PriRoAgg-RFA: Uses a geometric median approach to defend against model poisoning.

Datasets

The researchers evaluated their protocols using two standard computer vision benchmarks:

  • Fashion-MNIST: A dataset of 70,000 grayscale images of clothing items.
  • CIFAR-10: A dataset of 60,000 color images across 10 classes (airplanes, cars, birds, etc.).

They tested these against a gauntlet of attacks, including Gaussian noise, update scaling, "Min-Max" poisoning, and sophisticated backdoor triggers (like Trojan and IBA attacks).

Key Insights and Ideas

  • The "Minimum Leakage" Philosophy: The core "aha!" moment is realizing that robust algorithms don't actually need to see individual updates. They just need specific "aggregated statistics." By revealing only these sums, you get the benefits of outlier detection without the privacy cost.
  • Efficiency through Partitioning: By using LCC to partition model updates into $K$ pieces, the communication overhead for users is reduced by a factor of $K$. This makes the protocol much more scalable for large deep-learning models.
  • Repeated Pattern Optimization: The authors noticed that model updates apply the same math to every coordinate. They optimized their zero-knowledge proofs to take advantage of this "repeated pattern," drastically cutting down the computational cost of verifiability.

Key Results

  • Superior Robustness: In a "Min-Max" poisoning attack on CIFAR-10, PriRoAgg maintained 79.2% accuracy, while other "secure" protocols like RoFL and EIFFeL collapsed to as low as 10-18% accuracy.
  • Backdoor Immunity: Against the "Trojan A" backdoor attack on Fashion-MNIST, PriRoAgg achieved an Attack Success Rate (ASR) of 0.0%, effectively neutralizing the attack while maintaining high main-task accuracy.
  • Scalability: As the number of users ($N$) and the dimension of the model ($d$) grow, PriRoAgg's communication and computation costs scale much more favorably than baselines like EIFFeL. For large $N$, the use of LCC makes the "per-iteration" overhead significantly lower.

Limitations and Future Work

  • Semi-Honest Server Assumption: The current security proof assumes the server is "honest-but-curious" (it follows the protocol but tries to learn extra info). While the authors note that their protocol is more resilient than others, a truly "malicious" server that actively deviates from the protocol remains a challenge for future research.
  • Algorithm Constraints: The framework currently supports robust algorithms that can be expressed as an "aggregate of local functions." While this covers many state-of-the-art defenses, it may not cover every possible robust aggregator.
  • Multi-Server Potential: The authors suggest that extending this work to non-colluding multi-server scenarios could further strengthen privacy guarantees.

Summary updated

Abstract

Federated learning (FL), as a promising machine learning paradigm for large-scale distributed data, faces two security challenges of privacy and robustness: the transmitted model updates potentially leak sensitive user information, and the lack of central control over local model updates leaves the global model susceptible to malicious attacks. Current solutions attempting to address both problems under the one-server FL setting fall short in the following aspects: 1) design for simple validity checks that are insufficient against advanced attacks (e.g., checking norm of individual update); and 2) have partial privacy leakage for more complicated robust aggregation algorithms (e.g., distances between model updates are leaked for multi-Krum). In this work, we formalize a novel security notion of aggregated privacy that characterizes the minimum amount of user information, in the form of aggregated statistics of users’ updates, that is necessary to be revealed to accomplish more advanced robust aggregation. We develop a general framework PriRoAgg, utilizing Lagrange coded computing and distributed zero-knowledge proof, to execute a wide range of robust aggregation algorithms while satisfying aggregated privacy. As concrete instantiations of PriRoAgg, we construct two secure and robust protocols based on state-of-the-art robust algorithms, for which we provide full theoretical analyses on security and complexity. Extensive experiments are conducted for these protocols, demonstrating their robustness against various model integrity attacks, and their efficiency advantages over baselines.
BibTeX
@article{hou2025priroagg,
  author = {Sizai Hou and Songze Li and Tayyebeh Jahani-Nezhad and Giuseppe Caire},
  title = {{PriRoAgg: Achieving Robust Model Aggregation With Minimum Privacy Leakage for Federated Learning}},
  journal = {IEEE Trans. Inf. Forensics Secur.},
  year = {2025},
  url = {https://openreview.net/forum?id=p4zLMdKhZ8},
}

Loading related content...

Figure